How a sitting is read, and what the reading is worth

An epoch the transform could not turn into band powers has nothing to score. This is arithmetic rather than judgement, and it is checked before any state is looked at.

0.5–4 Hz is never empty in scalp EEG. Across 140 epochs of real human recording the lowest relative delta observed is 0.108, while 568 of 615 epochs stored by an earlier, broken signal chain report exactly 0.000000. A band that is exactly zero is a broken integration, not a quiet brain.

The rule is deliberately the only one in this gate. “Delta is high” says nothing — real awake EEG at these sites has a median relative delta of 0.605 and reaches 0.92 — so the classifier does not score delta at all. It reads it for this one thing it can honestly say.

Each state is a Gaussian in five coordinates, so an epoch can be asked how far it sits from the nearest one in units of that state’s own spread. Past a point, the honest answer is that it is not that state.

The threshold is derived, not chosen. It is the 99.9th percentile of the scoring statistic under the classifier’s own model, so at most one genuine epoch in a thousand is declined on distance — a false-refusal rate stated in advance rather than discovered afterwards. Change one of the five weights and the threshold moves with it.

Every channel is still screened on its own — for a flat trace, for saturation against the amplifier’s limit, for a frozen stream, for a slow drift that has swamped everything else, and for an amplitude no scalp produces. None of that detection changed. What changed is what is done with it.

The verdict, the list of failing checks and the signal-quality figure are all still computed, stored and shown. They now discount the confidence rather than replace the reading, because the screen is per-channel and the refusal was per-epoch, and because a caveat is actionable where a blank panel is not. On four channels a clean epoch is untouched, one dead electrode costs about a tenth of the claim, and a wholly unusable epoch costs four tenths.

30–50 Hz power sits on top of the frontalis and temporalis muscles, whose own peak falls inside the band, so high gamma is usually a clenched jaw rather than cortex. That physiology is real. The threshold that acted on it was not: it was the 99th percentile of a laboratory recording made with wet electrodes and a clinical amplifier, and this product runs on a dry band worn on a real head, where the band is never clean at any level.

Published comparisons of consumer and research hardware do not even agree on the direction of the gamma bias. A number that unreliable may lower a confidence; it should not be allowed to withhold a reading, which is the strongest thing this system can say. Elevated gamma now discounts every state instead, on a ramp anchored to the profile table’s own measured means.

The flag this replaced, and what it was really measuring.

Accepted epochs carry a ceiling. At a perfect fit it is 0.84, and it falls to 0.50 at the distance where the reading would be refused outright — the point where, by construction, “this state” and “nothing I recognise” are a coin flip.

The 0.84 is borrowed rather than measured, and that is stated because it matters. This judgement has no ground truth, so its accuracy cannot be tested; what can be quoted is how well trained experts agree with each other on the nearest comparable task, scoring graded drowsiness from EEG. Their agreement on the easiest cases in that literature works out at 0.84 across five categories, and it collapses to chance in the grey zone. Four seconds of four-channel dry-electrode EEG is not entitled to claim more than the best two humans manage on an easier question.

Niruddha is capped at 0.50 whatever the fit, by design rather than by measurement. It is the deepest state in the vocabulary and this classifier is stateless — it cannot see whether the neighbouring epochs agree — so a single four-second window is not allowed to claim it confidently. Earning the rest of that claim needs a session-level layer, and that layer does not exist yet.

Neither discount ever reaches zero. Both bottom out at the ratio of the two anchors above, 0.50 / 0.84, because an epoch read off nothing usable and an epoch sitting exactly on the refusal boundary are the same statement — the evidence has run out — and it would be arbitrary for them to disagree.

What the ceiling and the discounts take is left visibly unclaimed rather than redistributed, so the five states on screen do not sum to one. A sitting has long stretches where the signal genuinely does not separate two states, and naming one anyway would be the wrong kind of confident. A system that always has an answer is not more capable, only less informative about which answers to believe.

An earlier version of this argument, published here and elsewhere on the site, said the reading came back indeterminate when the margin between the top two states fell below 0.10. That rule is no longer the refusal path. It survives as a display flag and nothing more, and it is recorded as a correction rather than quietly dropped.

Where the distance threshold comes from, in full — and what the classifier this replaced got wrong.

Alpha power is averaged across the two right-hand channels and the two left-hand ones, and the reading is the log of their ratio. Below −0.15 Delve reports Iḍā, above +0.15 Piṅgalā, and the band between them Suṣumnā. The axis is a binary; the reading is not.

The bhūmi classifier does not consume this figure at all. Between- subject variance in frontal alpha asymmetry runs three to ten times the between-condition effect, so on a single session it is noise, and a number that noisy is not allowed to move which state is named. It is reported as svara, and it informs the guṇa balance, and that is its whole remit.

Everything above is about where the figure comes from. Whether a left-right alpha difference is the thing the texts call Iḍā and Piṅgalā is a separate question, and the honest answer is that the chain it rests on has not replicated. Frohlich et al. 2025 (PLoS ONE, high-density EEG, N = 19) found no significant lateralisation between left and right unilateral nostril breathing: t = 1.00, p = 0.33. Earlier work reporting the coupling has failed replication elsewhere, ipsilaterally or not at all.

Delve keeps reporting svara because it is the vocabulary the practice is taught in, and because the figure underneath it is a real measurement of something. It is not offered as evidence that the tradition’s account of that something is correct.

The failure mode worth stating: when both hemispheres fall under the extractor’s floor, the log ratio comes out at exactly zero — which looks like perfect equilibrium and is in fact an absent measurement. It travels with a validity flag for that reason, and the readings that depend on it are gated on the flag rather than on the value.

Exactly thirty-seven fields per epoch in the export, from the state and its confidence down to the raw band shares and the sensors underneath them. Nothing here is a summary of something else: these are the numbers the reading was computed from.

Several of those fields are measures of how organised the four seconds were rather than how much power sat in a band. They are what the corroboration layer reads when it agrees or dissents with the state above it.

TXT is a narrative report laid out for hand-annotation, so a teacher can mark up a sitting the way they always have. CSV and JSON carry the same source data for anyone who would rather plot it themselves.

A field that was not measured is written as an empty cell rather than a zero, which is the difference between “no reading” and “a reading of nothing” — and the difference between a plot with a gap in it and a plot with a lie in it.